Skip to content

Model frame when LHS is cbind (PR#14189)

3 messages · Brian Ripley, Arni Magnusson

Original
Arni Magnusson
# 
The model frame shows the response and predictors in a data frame with
nicely labelled columns:

  fm <- lm(wt~qsec+log(hp)+sqrt(disp), data=mtcars)
  model.frame(fm)  # ok

When the left hand side consists of more than one response, those response
variables still look good, inside a matrix:

  fm <- lm(cbind(qsec,hp,disp)~wt, data=mtcars)
  model.frame(fm)[[1]]  # ok

A problem arises when some of the response variables are transformed:

  fm <- lm(cbind(qsec,log(hp),sqrt(disp))~wt, data=mtcars)
  model.frame(fm)[[1]]  # ugh, missing column names

The model frame is useful for many things, even more so when all column
names are legible. Therefore I propose adding two new lines to
model.frame.default() between lines 371 and 372 in
R-patched_2010-01-12/src/library/stats/R/models.R:

    varnames <- sapply(vars, deparse, width.cutoff = 500)[-1L]
    variables <- eval(predvars, data, env)

NEW if (is.matrix(variables[[1L]]))
NEW     colnames(variables[[1L]]) <- as.character(formula[[2L]])[-1L]

    if (is.null(rownames) && (resp <- attr(formula, "response")) >
        0L) {

With this fix, the above example returns legible column names:

  fm <- lm(cbind(qsec,log(hp),sqrt(disp))~wt, data=mtcars)
  model.frame(fm)[[1]]  # nice column names

I hope the R development team can either commit this fix or improve it.

Thanks,

Arni
2 days later
Brian Ripley
# 
A few points.

0) This seems a Wishlist item, but it does not say so (see the section 
on BUGS in the FAQ).

1) A formula does not need to have an lhs, and it is an assumption 
that the response is the first element of 'variables' (an assumption 
not made a couple of lines later when 'resp' is used).

2) I don't think this is the best way to get names.  If I do

fm <- lm(cbind(a=qsec,b=log(hp),sqrt(disp))~wt, data=mtcars)

I want a and b as names, but that is not what your code gives. And if 
I do
[,1]     [,2]      [,3]

You've lost the names that the current code gives.

The logic is that if you use a lhs which is a matrix with column 
names, then those names are used.  If (as you did), you use one with 
empty column names, that is what you get in the model frame.  This 
seems much more in the spirit of R than second-guessing that the 
author actually meant to give column names and create them, let alone 
renaming the columns to be different than the names supplied.

3) It looks to me as if you wanted

cbind(qsec, log(hp), sqrt(disp), deparse.level=2)

but that does not give names (despite the description).  And that is I 
think a bug that can easily be changed.  That way we can fulfil yoour 
wish without breaking other people's code.
On Tue, 19 Jan 2010, arnima at hafro.is wrote:

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Arni Magnusson
    

    

      
      #
    

  


  

      
Thank you Prof. Ripley, for examining this issue. I have two more 
questions on this topic, if I may.

(1) Truncated column names

With your explanations I can see that the problem of missing column names 
originates in cbind() and the 'deparse.level' bug we have just discovered. 
I had tried different 'deparse.level' values, only to see that it didn't 
solve my problem of missing column names.

   attach(mtcars)
   cbind(qsec,log(hp),sqrt(disp), deparse.level=0)  # no column names
   cbind(qsec,log(hp),sqrt(disp), deparse.level=1)  # qsec only
   cbind(qsec,log(hp),sqrt(disp), deparse.level=2)  # no column names
   cbind(qsec=qsec,log(hp),sqrt(disp), deparse.level=2)  # works!
   cbind(qsec=qsec,log(hp),sqrt(abs(disp)), deparse.level=2)  # hmm...

Now a new question arises. The last line generates these truncated column 
names

   "qsec"  "log(hp)"  "sqrt(abs(d..."

where the dots are not mine, but something that R decided to do, 
presumably to keep the column names no longer than 13 characters. I would 
prefer to retain the full column names, like this,

   as.matrix(data.frame(qsec,log(hp),sqrt(abs(disp)), check.names=FALSE))

where the column names are

   "qsec"  "log(hp)"  "sqrt(abs(disp))"

Is there some reason why cbind() should truncate column names? Matrices 
have no problems with very long column names.


(2) Changing the default 'deparse.level' to 2

Furthermore, since many users appreciate the compact model formula syntax 
in R, it would be great if the formula

   cbind(qsec, log(hp), sqrt(disp)) ~ wt

would result in a model frame with full column names, without sacrificing 
legibility by adding deparse.level=2 in between the variable names. The 
simplest way to achieve this would be by changing the default value of 
'deparse.level' to 2 in cbind() and probably rbind().

Am I missing some important cases where functions/users rely on some of 
the column names to be missing, as generated by deparse.level=1? And if 
so, do these cases outweigh the benefits of clean and compact formula 
syntax when modelling?


Many thanks,

Arni

        
On Thu, 21 Jan 2010, Prof Brian Ripley wrote: